HAL Id: jpa-00246652
https://hal.archives-ouvertes.fr/jpa-00246652
Submitted on 1 Jan 1992
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Weak ergodicity breaking and aging in disordered systems
J. Bouchaud
To cite this version:
J. Bouchaud. Weak ergodicity breaking and aging in disordered systems. Journal de Physique I, EDP
Sciences, 1992, 2 (9), pp.1705-1713. �10.1051/jp1:1992238�. �jpa-00246652�
Classification
Physics
Abstracts75.40 05.40 64.70
Short Communication
Weak ergodicity breaking and aging in disordered systems
J. P. Bouchaud
Service de
Physique
de l'Etat Condensd,CEA-Saclay,
91191Gif-sur-Yvette Cedex, France(Received
25May
J992,accepted
infinal form
J9June J992)Abstract. We present a
phenomenological
model for thedynamics
of disordered(complex)
systems. Wepostulate
that the lifetimes of the many metastable states are distributedaccording
to a broad, power lawprobability
distribution. We show thataging
occurs in this model when the average lifetime is infmite. Asimple hypothesis
leads to a new functional fornl for the relaxation which is in remarkable agreement withspin-glass experiments
overnearly
five decades in time.In
spite
of fifteen years ofdispute,
thetheory
ofequilibrium spin-glasses
is notyet
settled[1, 2, 3]. Dynamical
effects are,however, likely
to be the dominant aspect inexperiments.
One of the most
striking
aspects of thedynamics
ofspin-glasses
in their low temperaturephase
is theaging phenomenon-
a ratherpeculiar
and awkward feature from thethermodynamics point
of view : the relaxation of a systemdepends
on itshistory.
Moreprecisely,
if a system is field-cooled below itsspin-glass temperature,
themagnetization
relaxation
depends
on theWaiting
time t~ between thequench
and the switch off of themagnetic
field[4, 5, 6].
Similar effects are observed on the viscoelasticproperties
ofpolymer
melts
[7], magnetic properties
of HTCsuperconductors [8]
and morerecently
on therelaxation after a heat
pulse
inCharge Density
Wave systems[9]. Analytical
fits of themagnetization
relaxation as a function of time have beenproposed.
In[6],
it isproposed
thatthe initial
«stationary»
part of the relaxation is apower-law
With a small(negative)
exponent. For times
longer
than theWaiting time,
relaxation is Well fittedby
a « stretched »exponential decay, provided
that an effective time is introduced[6].
In[4, 5], however,
astretched
exponential
of the «real» time was found forrelatively
short times.Many
phenomenological
theories have been devised to account for the stretchedexponential decay lo-14],
and for the slow part andaging [6, 15-18].
We however feel that the basic mechanismunderlying aging
has not beenfully appreciated (see
however[17]) although
itis,
in ouropinion,
one of the constitutiveproperties
ofspin-glasses.
The aim of this note is tosuggest
thataging
is related toergodicity breaking
which has, in these systems, apeculiar
andperhaps unexpected meaning.
We find on somesimple
models that-according
to ourdefinition,
see below « Weak »ergodicity breaking
occurs in thespin-glass phase,
definedas the
phase
Where the Edwards-Anderson parameter is non-zero. We see however no reasonWhy
the two should be linked ingeneral,
and suggest that the appearance ofaging
could be taken as anoperational
definition of thespin-glass
transition.1706 JOURNAL DE PHYSIQUE I N° 9
We
reproduce
both the short time(t«t~)
and thelong
time(t»t~) parts
of themagnetization decay
underquite
reasonablehypotheses.
The
analytical
form of thisdecay
is however different from all thosepreviously proposed.
We find in
particular
that :m(t)
= mo expi- y/(I x) (t/t~)~ ~~i
0 w x wfor t « t~ and a power law
decay
m(t)
=
(tit~)-Y
for t » t~. Note that the two
regimes
arestrongly connected,
since the above exponent is related to theparameters describing
the initialpart
of the relaxation. This feature should be astringent
test of ourtheory.
As We shall see, ourtheory
is in remarkable agreement With both the « short » timeexperiments
of[4. 5],
and the verylong
time data of[6].
It is
Widely accepted
that the energylandscape
of a finite disordered system isextremely rough,
With many local minimacorresponding
to metastableconfigurations,
Which We shall callloosely
« states ». Since the energylandscape
isrough,
these local minima are surroundedby
ratherhigh
energy barriers. We thusexpect
that these states act as « traps » Whichget
hold of thesystem during
a certain time r.What is the distribution of these
trapping
times~(r)
? Asimple picture
for the energylandscape
is thefollowing (Fig. I,
see also[11])
: there exists a «percolation
» energy levelfo
below which« states » are disconnected ;
fo
is the minimal energyrequired
to «hop
»between any two states. In other
words,
we assume that thedynamics
benveen the traps is very fast and that theprobability
to find thesystem
between two metastable states isnegligible.
r
(
a
r
b)
AGINGio
r
Fig.
I. a) Schematic view of the energylandscape
: holes are drilled below the reference energyfo
which is the minimal energy needed to go from one metastable state to another. Note that thisdrawing
is one dimensional : inreality
mountains ofheight
»f~
also exist between different states. b) When the extemal field is cut, thelandscape acquires
a non-zeroslope
which drives the system towards zeromagnetization
states.In this
picture,
the energy barrier AE is thusequal
to thedepth
of the trap,fo f.
In thespin glass phase
of both the RandomEnergy
Model(REM)
and of the SKmodel,
the distribution of very low f's isexponential [1, 23, III
:P
~f )
= NIT exp x
( f fo)lT (1)
where x is a temperature
dependent
number between 0 andI, fo
is the reference level above Which levelsproliferate
and N is a constant. In theREM,
x=
T/T~
Which is connected to the fact that the free energylandscape
istemperature independent.
This is not the case in the SKmodel,
Wherex(T)
is non-trivial[I].
Itsshape
is Wellapproximated by
the «one-step
»replica breaking
scheme[5].
Assuming
that AE=fo- f,
and thatr =
roexp(AE/T),
onefinds, using (I)
and~
(T)
dr=
P
( f ) df
:~(y)
~yx y- (I +x)(2)
0
for
large
r.[w
is a normalizationconstant].
Such a form is also obtained in many othermodels- for
example
the one-dimensional «random force model»II?, 20-22],
withx =
TF/« where F is the average force and wits variance.
It may be that for real
spin-glasses
theprobability
distribution of low free energy statesdecays
moreslowly
forlarge negative f,
forexample
as a power lawP
(f )
=
fi(- f )~
~~ +'~(3)
with ( ~ 0. In this case, ~
(r)
reads :~
(r )
=
air
jog rir~j-
(' + <)(4)
where 8 is a constant. This will lead to
logarithmic,
asopposed
to powerlaw, decay. (4)
wouldprobably
conform more to the ideas of Fisher and Huse[15].
The most
interesting property
of these distribution laws is the fact thatIT)
w=
drr~(r) diverges,
when xwl forequation (2)
and for all finite(
foro
equation (4).
Hence,
the time needed toexplore
an infinite system is infinite. This is however a non-conventional scenario for
ergodicity breaking,
since thephase
space is apriori
not broken intomutually
inaccessibleregions
in which localequilibrium
may be achieved. We shall call this situation « weak »ergodicity breaking.
It may be that« true »
ergodicity breaking
also occurs, I.e. the existence of many « pure states » between which infinite barrier stand. We shallcarefully
avoid this issue(the
reader is referred to[19]
for a recentexperimental discussion),
since thedynamics
isby
definition restricted toonly
one pure state, and for which the concept of « weak »ergodicity breaking
is relevant.Let us now see how this feature is
deeply
connected to thepossibility
ofobserving aging.
Suppose
that the system understudy
can bedecomposed
into ~lLindependent subsystems.
These
subsystems
maybe,
forexample, weakly coupled polycristalline
«grains
» or clusters ofmagnetic
atomsthey
are not, in ourmind,
« domains »growing
with time as is assumed in references[15, 16] (although
such domains may indeed be present inside eachgrain).
In otherwords,
we assume that anexperimental sample
isalready
a collection of alarge
number of microspin-glasses,
and hence that the measuredquantities
are disorder averages. We claim thatperforming dynamical experiments
onmesoscopic samples,
where ~lL issmall,
wouldyield irreproducible
results.1708 JOURNAL DE PHYSIQUE I N° 9
Each
subsystem
I(
=
I,
~lL) can be in a certain number of « states »8(I )
= 8 which we shall suppose
roughly independent
of I. The total number of states is thus=
8~
Note that in the SKmodel,
8depends
ontemperature,
with inparticular 8(T~)
= I.The system is field cooled to a certain
temperature T~T~.
Since amagnetic
field H is on, the system evolves within a «pool
of states» which carry a net
magnetization
butcannot reach states with a different value of this
magnetization
: this would cost an extensive(free)
energy. The free energy distribution of these states is characterizedby
a certainx(T, H)
or((T, H) depending
on which of the two distributions(I)
or(3)
is moreadapted. [For convenience,
we shalladopt
in thefollowing
the distribution(I),
thecorresponding
results for distribution(3)
can be obtained from thosegiven
belowby defining
a time
dependent
effective parameter x(t) according
tox(t
=
(I
+~) log(log t/ro)/log(t/ro)
which goes to zero as t -
oo].
The
magnetic
field is left on a time t~.During
thistime,
thesubsystems
startexploring
theirphase
space. Thedeepest
trap available for agiven subsystem
I is characterizedby
atrapping
time
t(I)
such that :s
°~
dr ~
(r)
=1(5)
t(1)
Equation (5) simply
means that out of 8trials,
a value ofr
larger
orequal
tot(I )
has been encounteredroughly only
once.(A
morerigorous justification
can be found in e.g.[21]). Thus,
from(5), t(I )
= To
8"~.
Thislongest trapping
time of course fluctuates fromone
subsystem
to theother,
with a distributiongiven by £(t, 8)
such that£
(t, 8)
=
r(
Sit + ~(6)
for t »
8~'i [The
maximum relaxation time available in the whole system isthus, similarly, t(~lL)
= 8~'~~~.]
Let us now estimate the order of
magnitude
of thelongest trapping
time r~~actually
encountered
during
a time t~, when r~~~ is much smaller thant(I).
If N(t~)
denotes the number of different « states» visited
by
thesubsystem
after t~, r~~~ is obtainedthrough
:N
(tw ) °~
dr ~
(r )
>
(7a)
lr~
t~ =
£ rj
m N
(t~)
dr r~(r ) (7b)
j i. N(iw) 0
which
gives,
for x~ l and for all finite (, r~~
= t~. This is the
key point
in ourpicture
: thedeepest
state encounteredtraps
the systemduring
a time which iscomparable
to the overallwaiting
time.Schematically,
states such that r «t~
arethus,
for x ~l, ergodically probed,
whereas those for which r » t~ are
barely
visited. This is not so for x~ l : from
(7a, b)
onefinds that the
typical
value of r~~~ isr/ "~(t~)~'~
« t~.We are now in a
position
to describe the observed relaxation lawsquantitatively.
When themagnetic
field is switchedoff,
the free energy of the metastable states areuniformly
shiftedby
an amount + MH
(Fig. lb).
Thesubsystems
will evolveindependently
towards statescarrying
no
magnetization,
whichrequires
a collectivechange
of a finite fraction of thespins.
The system must leave the metastable state that it had entered when the field was on.However,
the system may also continue toprobe deeper
anddeeper
wells whilerelaxing;
its age t~ is thus timedependent
with :t~(t)
= t~ + t. This feature was
recognized
as a crucialingredient
for a consistentanalysis
of theexperimental
data[6].
Two cases must then be
distinguished
:A)
Shortwaiting
times :Suppose
first that thewaiting
time t~ is much shorter thanTog "§
definedby equation (5).
In thisregime,
thephase
space iseffectively
infinite. Theprobability
P(r,
t~)
to find agiven subsystem
in a state characterizedby
r caneasily
be shown to beequal
tor(r/t~ ) r~(r),
with
r(u)=
I for small u, anddecaying
to zero for u»I. The detailed form ofr(u)
woulddepend
on the « geometry » ofphase
space I.e. the relativepositions
andinterconnections between the different metastable states. In the
simple
case where all states areequally accessible,
one findsr(u)
= I/u for u » I.
The very curious feature in the x
~ l
phase
is that r~
r is apriori
notnormalisable,
sinceIT )
is infinite. P(r,
t~)
can thusonly
be normalised when thewaiting
time t~ isfinite,
and then reads :P
(r,
t~)
=
Ar
(r/t~ ) (t~
)~'/r~ (8)
lw
with A
=
dur(u)/u~ [=
III -x if one takes asimple step
forr(u)].
Note that theo
microscopic
time To hasentirely disappeared
fromequation (8)
This would not be the case for x ~ l : in this case, onesimply
finds P(r,
t~)
cc r ~(r ),
which in factcorresponds
to theergodic
result(indeed, r~(r)
isproportional
to the Boltzmannweight).
Assuming
asimple exponential decay
fromsingle
states('),
we thusfind,
for themagnetization
relaxation rate :lm
dm/m
=
pi I/rj
dt= dt
lp/Zj
dr I/r P(r,
t +t~) exp(- t/r) (9)
o
lw
where Z
=
dr P
(r,
t + t~ exp(- t/r).
p is theprobability
for themagnetization
of theo
subsystem
to relax whenleaving
a r-trap, rather than continueaging
amongmagnetized
states. p thus measures « how easy » it is to find states with small
overlaps [24].
Aging during
relaxation is taken into accountby
the appearance oft + t~ in the aboveexpression.
Formula(9)
iseasily
transformed into :dm/m
= t
-~(t
+ t~)~~
' G(tit
+ t~)
dt(10)
where G
(u)
is anotherslowly varying function,
related tor(u),
such that G(O)
=
pr(x)/A.
Note that
equation (10)
is close inspirit, although
notequivalent,
to the treatment of references[6, 7],
where adescription
ofaging during
relaxation wasproposed.
In thefollowing
we shallneglect
the variations ofG,
andstudy (10)
in the short andlong
timeregimes.
For t « t~ one finds
m(t)
=moexpi- y/(I -x) (t/t~ )' ~~i
=
mo(I y/(I -x) (t/t~
)~ ~~ +) (II)
where y=
pr(x)/A.
For t » t~,aging during
relaxation becomes thepredominant
feature and we find :m
(t)
=
(tit~ )-
Y(12)
(1) Note however that
our central result, equation (10), is
unchanged
up to a redefinition of G if exp(- t/r) isreplaced by
any function of t/r decaying faster than t~1710 JOURNAL DE PHYSIQUE I N° 9
(assuming
thatG(I
= G(O)).
Three remarks must be made :I)
ourtheory reproduces
theubiquitous
« stretched »exponential decay,
which is seen to beonly
valid at short times t « t~.One should also note that based on the
exponential
distribution offree-energies (Eq. (I)),
De Dominicis et al.[I II
haveproposed
adynamical
model for which a stretchedexponential decay
is found forlong times,
with an exponent x(while
we find I -x, seeEq. (ll)).
However,
ourapproach
is very different since we assume that thehopping
rate from one state to another is determinedby
the «starting
state » and not, as in[I ii, by
the « destination »state. Furthermore, no
aging
effects were found in this model.ii) (I I)
and(12)
show that the two limits(short
andlong times)
areintimately
connected.iii)
Note that in thisregime,
m(t,
t~)
is a function of the ratiot/t~ only.
Inparticular,
we find that the short timedecay,
for t « t~, is notstationary.
This is at variance with thetheory
of references
[15, 16],
where theequilibrium (stationary) dynamics
ispredicted
fort«t~_ In our
model,
this isonly
the case when t~» To8"~ (see below).
Moreoverm(t~,
t~)
is still of order of mo.We have fitted some of the
experiments
of[6]
with the function obtained from theintegration
of(10),
with a constant G. This leaves threefitting
parameters: mo, x, and y.(Note
however that theshape
of the curve isonly parametrized by x).
As can be seen from
figure 2,
the agreement is remarkable in view of the broad scale of times which isprobed
: t varies from 0.I to 3 500min,
and thewaiting
time t~ isequal
to 31.3 min. For T= 0.6
T~,
we find x= 0.76 and y
= 0.062 ;
using
A=
I/(I x),
thiscorresponds
to p of the order of 1/3. Note inparticular
that analgebraic
distribution of lifetimes(Eq. (3))
appears to be moreappropriate
than thelogarithmic distribution, equation (4).
T=10K, x=0.7645, y=0.0645, m(t=0)=0 567
O.6
o.5
O.4
O.3
O.Ol 0.I 10 loo 1000 1001
log(t[min.])
Fig.
2. Fit of one of theaging experiment
onCdcrj_~Ino,354 spin glass
[6]. Thewaiting
time is 31.3 min under 15 Gauss, and the temperature is10K=0.6T~.
The set of paranJeters used wasmo = 0.567 (in units of the field cooled
magnetization),
x=
0.76 and y
=
0.0645.
Other
temperatures
andwaiting
times areequally
wellreproduced by
ourtheory.
A more extensivepresentation
of thesefits,
and of the veryinteresting
field and temperaturedependence
of the parameters, will bepublished
elsewhere[25].
Notefinally
that theexperiments
of[5]
have beenanalyzed
in terms of a stretchedexponential decay.
This 15nicely explained by
ourtheory,
since in theseexperiments,
thedecay
time is less than orcomparable
to t~.
In order to
emphasize
thepeculiarity
of theregion
x~I,
one should note that forx~l,
where the distribution(8)
becomesnormalisable,
one findsm(t)=mo
[I (t/ro)
+ : themagnetisation
has thusdrop substantially
after amicroscopic
time To, and not after themacroscopic
time t~.B) Long waiting
times and«
interrupted aging
»Let us now consider the other limit t~ »
Tog
~'§ whichcorresponds
to the case where most systems have reachedequilibrium. Using (6),
we find that(8)
isreplaced by
:p(y) gyy~l+xj~/gI/~ j gl-16( )~-l/~x (j~)
" T0 ~ T0
for r «
Tog
~'~Hence,
for t «Tog ~'~«
t~, we find thatm
(t)
=
mo(i yi(i x) (t/ro
si/x)i
-x) (14)
which is
independent
of t~.For
ro8"~«t«t~,
one observes those
subsystems
such thatt(I)~t~,
for whichequilibrium
is reached. From thetheory
ofL£vy
sums(see
e.g.[21]),
one may show that thedeepest traps
of thesesubsystems,
for x~
l, entirely
dominates thepartition
function. In otherwords,
theprobability
tofind
thesubsystem
I in itsdeepest
state isoforder
I. The fractionm(t)
ofsubsystems
stillcarrying
their initialmagnetisation
is thus, forTog "~«
t « t~,m(t)/mo lm
= dr
(rj8/r~ +~j exp(- t/r)
=
8(ro/t)~
« l(15)
o
where we have used the
probability
distribution(6).
Henceaging essentially
stops(or
becomes
inobservable)
whenTog
"~=t~~~, beyond
which the finite size of thephase
space appears. «Interrupted aging
» has indeed been observed inCharge Density
Waves[9],
andcan be used to measure the «
complexity
» n of the system, which isnaturally
defined asn
=
Logs
=xLog (t~~~/ro).
Let usfinally
mention thataging
of lowfrequency
a-c-susceptibility
can be understoodusing
an extension ofequation (9)
:x
(w, t~)
=
°~
dr p
(r,
t +t~) i/(i
+iwr) (16)
where a
simple Debye
form is assumed for each state[25].
Inparticular,
we find thatx
"(w, t~)
=
4 (&
= w
t~)
with4
&)
= & for small it, and
4
&)
=
&~ ' for
large
&i, in agreement with
Koper
and Hilhorst[16].
We have thus
proposed
aphenomenological
model which allows us tounderstand,
on verygeneral grounds,
how thephenomenon
ofaging
sets in : thephase
space must posses alarge
number of metastable states, which abroad, scaleless,
distribution of lifetimes. When the average lifetime of these metastable statesdiverges,
all thephysical
observables are dominatedby
theproperties
of thedeepest
state that thesystem
was allowed toprobe
in its1712 JOURNAL DE PHYSIQUE I N° 9
quest for the « true »
equilibrium
state.Aging
stops either when the distribution of lifetimes is truncated and is a finite sizeeffect,
or when the distribution of lifetimesdecays
morerapidly
than
r~~
Amore formal
description
of our results could begiven, along
the lines of reference[11], using
a Masterequation
inphase
space.The most
important quantity
of ourtheory
is the index x, whichnaturally
appears in the SKor REM models of
spin-glasses.
xgives
information on the distribution offree-energies
in thespin-glass phase,
and the present model 15 asuggestion
for itsexperimental determination,
which would be of great interest. The behaviour of x and the
disappearance
ofaging
asT~
isapproached
determines the nature of thefreezing
at theglass point
: in the RandomEnergy
Model[23], x(T~)
=
I,
which means that thephase
space is broken into manymetastable states of
comparable
« size ». In the SKmodel,
on the otherhand, x(T~)
=(one valley dominance) [II
and the number of statesS(T-T~)
is tends to I :aging
is theninterrupted
formicroscopic
times.Let us
finally emphasize
that our model is not restricted tospin-glasses
and may thus be useful to understandaging
in other systems such aspolymer glasses [7]
andCharge density
waves
[9].
Acknowledgements
I want to thank Eric Vincent with whom I have had many crucial
discussions,
inparticular
onthe
experimental
data. This paper has beenstrongly
influencedby
these discussions.Important
remarks andclaryfing
comments are also due to M.Feigelman,
J.Hammann,
F. LeFloch,
M. Ocio inSaclay
and P.Cizeau,
B.Derrida,
D. S.Fisher,
A.Georges
andM. M£zard.
Finally,
aninteresting
discussion with R. Orbach must beacknowledged.
Part of this work has beenperformed
at Laboratoire dePhysique Statistique,
Ecole NormaleSup£rieure.
References
[Ii MtzARD M., PARISI G., VIRASORO M., « Spin Glass
Theory
& Beyond» (World Scientific,Singapore, 1986).
[2] BINDER K., YOUNG A. P., Rev. Mod. Phys. 58 (1986) 801.
[3] FISHER D. S., HusE D.,
Phys.
Rev. Lett. 56(1986)
1601Phys.
Rev. B 38 (1988) 373.[4] LUNDGREN L., SVEDLINDH P., NORDBLAD P., BECKMANN O.,
Phys.
Rev. Lett. 51(1983)
911 NORDBLAD P., LUNDGREN L., SVEDLINDH P., SANDLUND L.,Phys.
Rev. B 33 (1986) 645.[5] CHAMBERLIN R. V.,
Phys.
Rev. B 30(1984)
5393HOOGERBEETS R., Luo W. L., ORBACH R., Phys. Rev. B 34 (1986) 1719.
[6] ALBA M., OcIo M., HAMMANN M., Europhys. Lett. 2
(1986)
45 J. Phys. Lett. 46 (1985) L-1101, for a review, see M. Alba, J. Hammann, M. Ocio, Ph.Refregier,
J. Appt. Phys. 61 (1987) 3683 in « RecentProgress
in RandomMagnets
», D. H.Ryan
editor (McGillUniversity,
1992).[7] STRUIK L. C. E.,
Physical Aging
in AmorphousPolymers
and Other Materials (Elsevier, Houston, 1978).[8] RossEL C., in Relaxation in
Complex Systems
and relatedtopics
(Plenum Press, New York, 1990).[9] BILJAKOVIC K.. LASJAUNIAS J. C., MONCEAU P., LtvY F.,
Phys.
Rev. Lett. 67 (1991) 1902 ; BILJAKOVIC K., LASJAUNIAS J. C., MONCEAU P., Phys. Rev. Lett. 62 (1989) 1512.[10] PALMER R., STEIN D., ABRAHAMS E., ANDERSON P. W.,
Phys.
Rev. Lett. 54 (1984) 958see also D. Stein, « Chance & Matter », J. Souletie, R. Stora, J. Vannimenus Eds. (North Holland, 1986).
[1II DE DOMINICIS C., ORLAND H., LAINtE F,, J.
Phys.
Lett. 46 (1985) L-463.[12] MONTROLL E. W., BENDLER J, T,, J. Stat.
Phys.
34(1984)
129, [13] CAMPBELL I. A., J. Phys. Lett. 46 (1985) L-1159.[14] RAMMAL R,, J, Phys. 46 (1985) 1837
CASTAING B., SOULETIE J., J. Phys. J France 1 (1991) 403, [15] FISHER D, S,, HusE D.,
Phys.
Rev. B 38(1986)
386.[16] KOPER G, J. M,, HILHORST H. J,, J.
Phys.
France 49 (1988) 429 ;see also H,J, Hilhorst, in Fundamental Problems in Stat, Mech. VII, H. Van
Beijeren
Eds.(Elsevier, 1990).
[17] FEIGELMAN M. V., VINOKUR V., J.
Phys.
France 49(1988)
1731DOTSENKO V. S., FEIGELMAN M. V., VINOKUR V.,
Spin
Glasses and Related Problems, Soviet Scientific Reviews, vol. 15 (Harwood, 1989).[18] SIBANI P., HOFFMANN K.,
Phys.
Rev. Lett. 63 (1989) 2853 SIBANI P.,Phys.
Rev. B 35 (1987) 8572.[19] LEDERMAN M., ORBACH R., HAMMANN J., OCIO M., VINCENT E., Phys. Rev. B 44 (1991) 7403.
[20] BoucHAuD J. P., COMTET A., GEORGES A., LE DoussAL P., Ann. Phys. 201 (1990) 285.
[21J BoucHAuD J. P., GEORGES A., Phys.
Rep.
195 (1990) 127see also
Appendix
A of BoucHAuD J. P., OTT A., LANGEVIN D., URBACH W., J.Phys.
ii France 1 (1991) 1465.[22] DERRIDA B., POMEAU Y.,
Phys.
Rev. Lett. 48 (1981) 627.[23] DERRIDA B.,
Phys.
Rev. B 24(1981)
2613 ;DERRIDA B., TouLousE G., J.
Phys.
Lett. 46 (1985) L-223.[24] ORBACH R.
(unpublished)
argues that p can be related to the « branching ratio » of the ultrametric tree in Parisi's solution. Note that he also obtains analgebraic
decay of m (t) atlarge
times.[25] BoucHAuD J. P., VINCENT E., HAMMANN J., Ocio M., in